This paper describes the use of a pyramidal architecture in recognizing and obtaining coordinate information of the neurons in an image. The input is an image from an electron microscope of the neuropil segment of the larval corn borer.

The pyramidal structure of 10 levels (numbered 0–9). The input image is of 256 × 256 size. A 64 × 64 portion of it is presented as test data and is stored at level 6. The nodes at level 5 perform median filtering on a 3 × 3 child set of level 6. Then Prewitt’s edge detection operators (for 8 directions) are applied to each of the nodes at level 5, and the ones having the maximum weight are retained. These weights and the directions are assigned to the microedges so obtained.

On the next level (level 4) the edges are combined to obtain short curves and the weights of the short curves are computed as w = &Sgr;w_i &Sgr;w_j, where w_j is the weight of the microedge j (j=1, . . . ,8). Here also a 3 × 3 child set is used, and the sums are taken within this window.

The short curves are further combined to form long curves at level 3, again using a 3 × 3 window. Finally these are merged to form a whole nerve cell. If at this stage the system detects some unmatched pairs of curves, it looks for the missing curves during a second pass with a larger window size and lower threshold. If this also fails, the system still predicts the cells but with a lower confidence level.

The key features of this method are small windows, local operations, and highly parallel computations.

It is to be noted that increasing the window size increases overlapping and introduces redundancy. To some extent this makes the system insensitive to noise. The system also preserves key features like edges even at low resolutions. The results of running the system on a test image are presented in the paper.